There is a hum that follows you home from a data center. It lives in the ears the way other sounds don't. I remember it from the first server rooms I walked through in Melbourne—the stacked chassis of drives, the fans pressing air like a thousand small animals breathing, the low urgent whisper that says something here is working harder than you are. So when I read, in the middle of 2025, that SpaceX and Nvidia were building a data center in orbit, I did what anyone should do with a beautiful sentence. I checked whether it was true.
It wasn't. Not in the way the sentence suggested.
The report, published by Crypto Briefing, claimed the two giants were joining forces on an orbital data center—a node of AI compute floating in low Earth orbit, lashed to Starlink's laser mesh, powered by solar panels the size of tennis courts. The headline carried the weight of inevitability. The body carried almost nothing: no sources, no technical specifications, no timeline, no named executives, no budget figures. Just the skeleton of an announcement with the flesh missing. Weeks of independent reporting turned up no official confirmation from either SpaceX or Nvidia. What existed, according to industry whispers around June, was something softer: exploratory conversations about using Starlink's inter-satellite laser links as the communications backbone for a potential space-based data center. Exploration, not excavation. Discussion, not construction. The word "building" was doing the kind of heavy lifting that usually precedes a pulled muscle.
I know this shape. I have been chasing it through ledgers and whitepapers since 2017, when I was a junior security researcher in Melbourne, auditing an ERC-20 token called Project Etherium that promised decentralized cloud storage. The economic model was broken—I wrote a two-thousand-word exposé called The Architecture of Hope demonstrating how the tokenomics would collapse under their own weight. And yet the founder's rhetoric of digital sovereignty was so seductive, so beautifully voiced, that early adopters shared my analysis not as a warning but as a form of appreciation. They liked the dream more than the breaking. That was my education: technical correctness is secondary to narrative cohesion in driving market sentiment. The story is the product. The code is just the packaging.
So let me be precise about what this orbital data center story is and what it is not. It is not a confirmed project. It is not a verifiable milestone. It is not even a signed memorandum of understanding, as far as public records show. What it is, is a narrative event—a signal fired into a market already trembling with compute anxiety. My job, as I see it, is to trace the ghost in the whitepaper's code and figure out who built the machine that makes the ghost visible.
Context: The Long Road to Orbit
The concept of a data center in space is not new. It has haunted the industry for decades, a recurring dream of engineers who looked at the Earth's surface and saw limits: land, water, power, politics. In the early 2000s, NASA-funded studies examined orbital server clusters for communications and sensing. In the 2010s, the idea resurfaced in conference slides as a thought experiment about latency and solar power. It never got past the thought because the physics never got past the physics. But the compute landscape of 2025 is different. AI training clusters have grown to consume hundreds of megawatts. Hyperscalers are fighting over nuclear power plant off-take agreements the way nations fight over oil fields. Data center leases are being signed years before construction completes. The anxiety is real, and into that anxiety, the orbital data center steps as a kind of savior figure—a white knight arriving in the silence of the void.
The two companies at the center of the rumor are the only two on Earth that could plausibly pull it off. SpaceX brings the launch monopoly: Falcon 9, Starship, and the largest low Earth orbit constellation ever built, with thousands of Starlink satellites already in operation, linked by inter-satellite laser connections that have reached ten gigabits per second per link. Nvidia brings the compute monopoly: more than ninety percent of the AI training market, a CUDA ecosystem that has become the operating system of modern artificial intelligence. Together, they would cover the full stack of what an orbital data center would need: transportation, communication, computation. It is a formidable pairing, and precisely because it is formidable, the rumor has legs. It doesn't have to be true to be believed. It just has to be plausible.
The rest of the field is a collection of earnest, underfunded attempts to plant flags before the giants arrive. Lumen Orbit, a startup founded in 2024, wants to put GPU clusters in orbit, with a first test satellite promised for 2025. The European ASCEND project, led by Thales Alenia Space, completed a feasibility study in 2023 and concluded that an economically viable orbital data center was possible only by 2036 at the earliest—and even that conclusion came with heavy caveats about power and cost. A handful of academic groups in Japan and Canada are studying radiation-hardened compute. None of them have launched a single GPU. None of them have actual customers. The entire sector rests on a foundation of PowerPoint slides and feasibility studies, which is exactly the kind of foundation on which narratives are built.
And here is a detail worth sitting with: a crypto media outlet broke this story, not TechCrunch, not The Verge, not Reuters. The readers of that outlet are mostly digital asset investors, people who have been trained by a decade of decentralized infrastructure narratives to see every piece of hardware in the sky as a node in a coming mesh of permissionless computation. The editors who chose to run this story knew exactly what they were doing. They were not reporting news. They were broadcasting a wavelength.
Core: The Physics That Refuses to Negotiate
Let us go to the vacuum, because the vacuum is where the story dies.
The first thing you must understand about a data center in orbit is that it would be the quietest data center ever built, and the quiet is the problem. On Earth, we cool our machines with air and water, moving heat away through convection and conduction. In a vacuum, there is no air. There is no water. There is only radiation: the slow bleeding of infrared energy through the void, governed by the Stefan-Boltzmann law, which says that radiative heat transfer scales with the fourth power of temperature. To radiate heat effectively, you must make the object very hot, or you must give it enormous surface area. An Nvidia H100 GPU, with a thermal design power of seven hundred watts, does not have enormous surface area. It has the footprint of a server card. Put it in a vacuum and it will cook itself unless you engineer an elaborate system of heat pipes, liquid ammonia loops, and radiator panels sized like sails. Every kilogram of that system is a kilogram that must be launched. Every kilogram that is launched costs money.
And then there is the power. Let me give you some numbers, based on my experience analyzing infrastructure economics across the crypto and AI worlds. A satellite in the one-thousand-kilogram class, with solar panels stretched like wings, might generate somewhere between ten and twenty kilowatts of electrical power. After you subtract the satellite's own operating needs—attitude control, communication, thermal management—you are left with perhaps five to ten kilowatts for computation. That is enough to run seven to fourteen H100 GPUs. A single ground-based AI server with eight GPUs has comparable compute. A modern training cluster has hundreds of thousands of GPUs. The scale gap between the orbital data center and the terrestrial data center is not a factor of two or ten. It is four or five orders of magnitude. The International Space Station, the largest structure ever placed in orbit, generates roughly one hundred twenty kilowatts—enough to power a small neighborhood, comically insufficient to power a training run.
The bandwidth story is no kinder. Starlink's inter-satellite laser links are genuinely impressive, a ten-gigabit laser mesh that has turned the constellation into a coherent network in the sky. But a ground data center's internal fabric operates in the hundreds of gigabits to terabytes per second range. NVLink and InfiniBand interconnect thousands of GPUs at speeds that make ten gigabits look like a telegraph. If you wanted to train a large language model across a distributed cluster of satellites, the communication overhead would dominate the training time so completely that the exercise would be mathematically absurd. The orbital data center is not a place for pre-training. It is, at best, a place for inference—the act of asking a trained model questions—and even then, the latency of a round trip through the sky adds twenty to forty milliseconds, which is fine for some workloads and fatal for others.
I want to pause on the economics, because economics are where narratives go to be tested. A one-thousand-kilogram satellite carrying ten GPUs, at a launch cost of roughly one hundred dollars per kilogram on a mature Starship, costs about ten million dollars just to put in orbit. That is a million dollars per GPU before you have bought the GPU, before you have built the satellite bus, before you have paid for radiation shielding, before you have developed the thermal systems, before you have insured the thing. A ground-based data center deploys a GPU for thirty to fifty thousand dollars, including the server, the power distribution, the cooling, the building. Even over a three-year operational life, the total cost of ownership for an orbital GPU is at least an order of magnitude higher. Ten times the cost. For what? For zero carbon? The launch itself, a single Falcon 9 mission, emits three hundred to five hundred tons of carbon dioxide; a Starship launch is in the thousands of tons. The zero-carbon-compute narrative collapses under the weight of the rocket fuel required to deliver the compute in the first place.
This is the paradox at the heart of the dream: the orbital data center is a solution to a problem that doesn't exist yet, at a price no one can afford, using technology that hasn't been invented. That is not a condemnation. It is a description of what early-stage infrastructure moonshots look like. But it tells you something about the nature of the claim being made. When I audited Project Etherium in 2017, I found the same structure: a beautiful answer to a question no one was asking, priced as if the physics had already been solved. The story was the product then, and the story is the product now.
So why are these two companies talking at all? Let me offer a more cynical and, I think, more accurate reading. Nvidia does not need a data center in orbit. Nvidia needs options. The ground-based expansion of AI compute is running into hard ceilings: power interconnection delays measured in years, physical space constraints in every major metro, and a growing backlash against the water and electricity consumed by data centers. In this environment, every possible source of compute becomes a hedge. Space is the ultimate hedge—the final frontier of somewhere else. It is a marginal capacity strategy dressed in the language of a revolution. For SpaceX, the motivation is equally unromantic. The company has built a transportation network and a communication network; the natural next step in its vertical integration story is computing. Launch plus low Earth orbit connectivity plus orbital compute equals a one-stop shop for space infrastructure. It is not a technical vision. It is a business plan, and it is a good one, because it turns Starlink from a commodity communication service into the bandwidth layer of an entirely new industry.
And then there is the layer that the public articles rarely mention, the layer of sand and gunmetal and classified briefings. A data center in orbit is a data center that never touches the ground. It does not transmit data through undersea cables that can be cut. It does not route traffic through foreign jurisdictions. It processes sensor data on the spot, in the silence, and returns only the results. For the United States Space Force, which has explicitly identified on-orbit computing as a critical capability, this is not science fiction; it is a doctrine in waiting. If SpaceX and Nvidia are exploring this path, you can assume that conversations with defense customers are already happening, and you can assume that the defense customers are willing to pay prices that make commercial launch costs look like pocket change. The economics I sketched above collapse when the customer is a government with a security requirement. For a defense agency, a ten-million-dollar launch for a data center that cannot be bombed is a bargain.
Let me also address the data sovereignty angle, because it is the most intellectually interesting part of this story and the most ignored. The theory runs like this: low Earth orbit is not sovereign territory. A data center floating there, operated by a private company, might theoretically escape the jurisdiction of any single nation's data protection regime. European companies subject to GDPR could process data in orbit without triggering cross-border transfer restrictions. Chinese firms could avoid certain data localization mandates. Multinationals could route sensitive analytics through the void. It is a beautiful theory, and it is almost certainly wrong, because the satellites themselves are subject to the jurisdiction of their launch state, and the companies operating them are subject to the laws of their incorporation. A data center in orbit does not escape law. It just moves the venue. The legal uncertainty is profound, and in that uncertainty, there is no compliance solution—there is only a compliance gamble. The value of the gamble, if it pays off, is enormous. The probability that it pays off, given how slowly international law moves, is low. But here is the thing about narratives: they don't need probability. They need possibility.
What would actually fly
If I were to design an orbital data center that made sense—and after the 2022 bear market, I developed a habit of asking exactly this kind of question in my series The Silence Between Candles, in which I interviewed engineers and miners and burned-out traders about what infrastructure we should build when the hype cools—I would not build a training center. I would build a sensing and inference edge node. The killer workload is satellite imagery. There are already hundreds of Earth observation satellites in orbit, generating terabytes of data every day, most of which is downlinked to ground stations, transmitted to data centers, processed, and then transmitted back up in the form of instructions. The round trip is wasteful. An orbital data center with a modest stack of GPUs could process that imagery in place, detect ships, track storms, identify deforestation, and downlink only the results. The compute requirements are modest. The latency requirements are forgiving. The bandwidth requirements are the reverse of training. This is the use case that survives contact with physics.
The same logic applies to the broader trend of on-orbit processing. As constellations grow—Starlink alone plans tens of thousands of satellites—the amount of data generated in orbit grows with them. At some point, the downlink bottleneck becomes the dominant constraint, and distributed processing becomes not a luxury but a necessity. Nvidia knows this. That is why, in my view, the real long-term product here is not a data center at all. It is a radiation-hardened, vacuum-cooled, low-power AI accelerator designed specifically for spacecraft. Think of it as the Orin chip that Nvidia builds for autonomous vehicles, but tuned for the radiation environment of low Earth orbit. If Nvidia develops such a chip, it will not need SpaceX to build a data center. It will sell the chip to every satellite manufacturer on Earth. That is the business model. The orbital data center is the demo. The chip is the product.
And this is where the infrastructure analysis becomes the most important lens of all. The bottleneck for orbital compute is not the GPU. It is the entire lifecycle of the system: the power generation, the thermal rejection, the radiation shielding, the bandwidth to the ground, the reliability under temperature swings of plus or minus one hundred degrees Celsius, the maintenance constraints of a machine that can never be touched again. Radiation doses in low Earth orbit run ten to fifty kilorads per year depending on altitude and shielding. That is enough to flip bits in memory and degrade silicon over time. A commercial GPU designed for the controlled climate of a server room is not ready for that environment. It needs hardening, and hardening costs performance. The engineering challenge is not building a data center in space. The engineering challenge is building a computer that survives space while computing at competitive speed.
There is a deeper infrastructural point here, and it is about the difference between centralization and distribution. The ground data center is a cathedral: massive, concentrated, protected by physical walls and armed guards. The orbital data center, as imagined, is a different creature entirely—a distributed mesh of compute nodes swimming through the void, each one a router, each one a processor, each one connected to the next by lasers. That is not a cathedral. That is a nervous system. And if you have spent any time in the crypto world, you will recognize the architecture: it is a decentralized physical infrastructure network, deployed in the ultimate jurisdiction. Whether you call it DePIN or you call it a starship, the shape is the same.
Contrarian: The Manufactured Emergency
This is where I want to introduce the contrarian lens, because the story of the orbital data center is not really about the orbital data center. It is about the anxiety that produced it, and about who benefits from keeping that anxiety alive. For years, venture capitalists and infrastructure startups have been telling us that liquidity fragmentation is a crisis—that the dispersion of users and assets across dozens of Layer 2 networks threatens the future of decentralized finance. The solution, conveniently, is always a new product that requires investment. But liquidity fragmentation, to the extent it exists, is a byproduct of the architecture that those same parties built, and it is survivable without the new product. We are watching the same manufactured emergency play out in AI infrastructure. The narrative says: compute scarcity is so acute, so existential, that we must consider launching GPUs into the heavens. The narrative's beneficiaries are the companies that own the launch vehicles, the GPUs, the data center real estate, the power producers. The anxiety is real. The solution space it opens is carefully curated.
Ask yourself who profits from a headline that says SpaceX and Nvidia are building an orbital data center. SpaceX profits, because it feeds the story that a transportation company is becoming a full-spectrum space infrastructure operator, which supports its valuation in private markets. Nvidia profits, because it feeds the story that the AI chip maker is not just a hardware vendor but the architect of compute across every possible environment, which supports the narrative premium in its stock price. The crypto outlet profits, because it reinforces the connection between decentralized infrastructure narratives and cutting-edge physical technology. And the rest of us? We are left with a longing, beautifully packaged, for a future that the physics does not yet support.
Alchemy in the age of open protocols: the story is the only thing that has actually launched. The engineering milestone—a GPU powered on in orbit, processing a real image and sending back a result—has not happened. Lumen Orbit's test satellite has not launched. ASCEND has no dedicated budget. SpaceX and Nvidia have made no official announcement. And yet the story is already in the world, doing work, changing weather. That is the true subject of this article. Not the data center. The machinery that produces the belief in the data center.
Let me also raise the uncomfortable question of what this story does to the global conversation about space governance. Low Earth orbit is a finite resource. It is already crowded with tens of thousands of tracked objects, and millions of untracked fragments. A constellation of large, power-hungry data center satellites would consume orbital slots and spectrum frequencies that others need. The wealthy West, led by a handful of corporations, would be effectively privatizing orbital real estate for compute. The space treaties that govern this domain were written when space was a Cold War theater, not a commercial compute farm. They are not equipped to handle questions like: who owns the orbital slot with the best solar exposure? Who has the right to radiate infrared heat into the void? Who is responsible when a ten-ton data center satellite collides with a defunct rocket stage and scatters thousands of fragments across the most useful orbital band? These are not hypotheticals. They are inevitabilities if the orbital data center becomes more than a story.
And then there is the militarization angle, which I have touched on but want to press further. In-orbit AI processing has a dual-use quality that cannot be ignored. The same compute that can classify ships for commercial shipping companies can classify warships for a navy. The same processing that can detect wildfires can detect missile launches. The line between civilian orbital compute and military orbital compute is not a line at all; it is a gradient, and it is already being walked by the United States Space Force, which has publicly identified on-orbit computing as essential to maintaining advantage. If the SpaceX-Nvidia project proceeds, other spacefaring nations will see it not as a commercial venture but as a strategic military capability. The response will be countermeasures: anti-satellite weapons, jamming, directed energy. The orbital data center will become a target. Anything that becomes a target becomes a weapon, whether it wanted to be or not.
The echo of a promise unkept
Here is what I believe, having spent a decade and a half watching narratives do their slow, alchemical work on markets and minds. The orbital data center is real the way a mirage is real: it exists as an image, vivid and precise, and it will shape behavior even if it never touches the ground. The image is doing three things right now regardless of whether a single GPU has left the atmosphere. First, it is legitimizing the compute scarcity story, reinforcing the idea that no amount of ground infrastructure will ever be enough, and thereby justifying the massive buildout of terrestrial data centers that is already underway. Second, it is seeding the concept of orbital jurisdiction, preparing the legal and political imagination for a future in which data processing happens beyond national borders. Third, it is giving cover to defense programs that want to put compute in space but prefer to ride the coattails of commercial glamour.
For the investor reading this, I have a simple framework, one I developed during the quiet months of 2022 when every headline was screaming sell and the only honest thing to do was to sit with the fear. Separate the milestone from the myth. The myth is the headline: SpaceX and Nvidia are building a data center in orbit. The milestone would be any of the following: an officially announced partnership, a test satellite manifest with a GPU payload, a successful in-orbit ignition of an AI accelerator, a first customer contract for orbital compute. None of those milestones have occurred. Until one of them occurs, treat this story as what it is: a signal of intent, a marketing event, a dream in search of a building permit.
And yet—and here is the part that keeps me from being purely cynical—the dream itself is not foolish. The longing for a form of computation that does not burden the earth, that does not consume rivers of water and forests of coal, that can process the world's data in the silence where no one can bomb it, that longing is real and it is not going away. The orbital data center, in its current form, is a premature embodiment of a legitimate desire. The heat problem, the power problem, the bandwidth problem, the cost problem: these are not refutations. They are design constraints. They are the walls of a maze that some future generation of engineers will navigate, perhaps with radically different technology than anything we can name today.
I think about this whenever I walk past a data center in the early morning, when the hum is almost liquid, when the building seems to breathe. The hum is the sound of a species trying to think. It is loud and hungry and it wants to be quiet someday. The orbital data center is an attempt to imagine what quiet looks like. The engineering is not there yet. The economics are not there yet. The governance is not there at all. But the imagination is there, and imagination, as I learned from a broken token called Project Etherium and a thousand other beautiful failures since, is the only fuel that matters.
So watch the sky. Not for a data center—you will not see one for years, if ever. Watch for the smaller things: a test satellite the size of a dishwasher, a glint of laser light between two points of metal, a brief press release from a company that says we are exploring a concept that may change everything. Treat those as the milestones they are. And remember, when the next headline arrives—when someone tells you that the future of compute is in orbit, or in the ocean, or in a ledger, or anywhere that sounds like freedom—that the story is being told for a reason, and the reason is not usually your enlightenment.
The vacuum will still be there, patient and silent, when the myth fades. The hum of the Earthbound data centers will continue, hungry and beautiful. And somewhere in a conference room, a company will be deciding whether to fund the next chapter of the story. I hope they fund the physics. I hope they fund the chip that is actually useful. I hope they fund the satellite that processes one image and sends it home. And I hope, most of all, that we all learn to measure the distance between the sentence "are building" and the quiet truth of what has actually been built. That distance, measured honestly, is the whole game.
